Methodologies for assessment of limit of detection and limit of identification using surface-enhanced Raman spectroscopy

Abstract

Here we rationally evaluate surface-enhanced Raman spectroscopy (SERS) substrates in terms of limit of detection (LOD), limit of identification (LOI) and dynamic range for ten common narcotic drug analytes. The drugs were amphetamine, cocaine, methadone, diazepam, methylphenidate, oxazepam, tramadol, morphine, buprenorphine and 6-monoacetylmorphine. A Raman microscope system was complemented with portable instrumentation, both in conjunction with commercial SERS substrates, and, by vibrational peak assignments, the functionality of substrates and pureness of samples was ensured. The dynamic range is explored qualitatively by concentration series measurements, where the Langmuir adsorption isotherm provided good fits. Moreover, an output fit parameter, the inverse of Langmuir constant, was found to roughly scale with LOD and can therefore be helpful in SERS substrate evaluations. Four different statistical methodologies were tested to estimate LOD: (i) a general formula to calculate a one-sided prediction interval for the mean value of blanks (LODB), (ii–iii) calculated from a one-sided prediction interval (at significance level 0.05) of a linear regression line, where the obtained limit of detection in the signal domain was sometimes outside the linear concentration range, which is why the corresponding concentration was calculated from (ii) a linear calibration curve (LODLR) and (iii) a non-linear calibration curve (LODNR), and (iv) using receiver operating characteristic (ROC) curves to estimate LODROC. Here, a new optimization approach was introduced for LODROC estimation, based on interpolation and thus better suited to handle a few data points spanning a large concentration range. LOI was assessed by discriminant analysis of partial least squares (PLS-DA) classification for seven of the drug compounds using PLS-DA, and the extracted LOIs were found to be higher than the LODs and were varying with respect to accuracy of the model which is strongly correlated to the probability of false positive detection that can be accepted.